OpenAI revealed Sora 2’s safety principles: from deepfakes to content labeling

OpenAI published a detailed description of the safety approach underlying Sora 2 — the new version of the video generation model — and the corresponding content creation application. The company emphasizes: security was not added on top of a finished product — it was built into the architecture from the very beginning of development. Sora 2 represents a new generation of OpenAI's video model, capable of generating realistic videos from text and visual prompts with substantially higher quality.

Simultaneously, OpenAI launched Sora as a standalone social platform oriented toward creative users: artists, directors, bloggers, and content creators. It is precisely the combination of a powerful model and an open public platform that, in the company's view, creates fundamentally new security challenges — ones the industry has not previously faced at this scale.

Protection is organized at multiple levels. The first is control at the model level itself. Sora 2 is trained to reject requests that violate acceptable use policy: creation of deepfakes of real people without their consent, sexualized content involving minors, materials promoting violence or spreading disinformation. This layer of protection is built directly into the model weights and activates before content is even generated. The second level comprises platform-side measures. The Sora app includes age verification, regional restrictions on certain types of content, complaint systems, and moderation tools that allow users to report violations.

Special attention deserves the work with content attribution. All videos created through Sora are marked using the C2PA standard (Coalition for Content Provenance and Authenticity) — a set of technical metadata that allows identifying material as AI-generated. This means that even after downloading and republishing a video, its origin can be established technically. The initiative is aimed at combating disinformation: editorial offices, platforms, and ordinary users will be able to see what was created by AI rather than captured by camera.

Another important element is external testing. Before public launch, OpenAI conducted large-scale red teaming: independent security researchers, filmmakers, and human rights organizations stress-tested the model, attempting to identify vulnerabilities and ways to circumvent restrictions. Their findings directly influenced the final configuration of the security system.

The company openly acknowledges: no single protective mechanism provides absolute guarantees. The bet is placed on layered protection — a combination of model limitations, platform rules, technical attribution, and moderation tools. Instead of seeking one perfect filter, OpenAI is building a system of mutually complementary barriers, each of which complicates abuse. Launching powerful video models into public access is always a compromise between creative potential and risks. The more realistic the synthesized content, the higher the stakes for society. The real test of OpenAI's approach will come not at the moment of release, but as Sora 2 begins to be used by millions of people in the most varied — including unpredictable — contexts.